Cognition's new AI cuts coding costs while rivaling top models
What's the story
Cognition has launched its new coding agent, SWE-2. The company claims that the new model is close to Anthropic's Fable 5.1 in real-world benchmarks but costs up to 70% less to run. SWE-2 scored 50% on the FrontierCode 1.1 Main benchmark, which assesses whether a human maintainer would merge an AI-written pull request. This is just shy of Anthropic's Fable 5.1, which scored 50.9%.
Training details
How SWE-2 was trained
SWE-2 is post-trained from Moonshot AI's Kimi K3, a 2.8 trillion parameter base model, using reinforcement learning across three effort levels: medium, high, and max.
Cognition claims it trained those effort levels in a single run, rather than building separate models for different tasks.
On the FrontierCode 1.1 Main benchmark, SWE-2's medium setting outperformed its predecessor SWE-1.7 while taking 58% fewer turns and costing 81% less on average.
Performance gaps
SWE-2 lags behind on more complex tasks
Despite its cost advantage, SWE-2 lags behind on more complex tasks.
On Terminal-Bench 4, a much harder agentic benchmark, it scored just 27.3%, far behind Fable 5.1's 55.8% and GPT-6 Astra's 57.9%.
This indicates that while SWE-2 is efficient for coding work where it can stay within a cleaner software-engineering lane, it still has a long way to go in more complex tasks compared to Anthropic and OpenAI models.
Strategic moves
Cognition acquires Poke, integrates SWE-2 into its product surfaces
Cognition has also acquired Poke, the only AI agent officially supported by Apple inside iMessage. The move is part of a strategic shift toward ambient consumer AI embedded in SMS rather than standalone apps.
SWE-2 has been integrated into Cognition's product surfaces, but the company hasn't published standalone model weights or a public per-token pricing sheet for it yet.